THE RELATIONSHIP BETWEEN TRAIN LENGTH AND ACCIDENT CAUSES AND RATES

Transcription

1 THE RELATIONSHIP BETWEEN TRAIN LENGTH AND ACCIDENT CAUSES AND RATES Darwin H. Schafer II Corresponding Author Graduate Research Assistant, Railroad Engineering Program Department of Civil and Environmental Engineering University of Illinois at Urbana-Champaign B-118 Newmark Civil Engineering Laboratory 205 N. Matthews Ave. Urbana, IL, USA Phone: (217) Fax: (217) Christopher P.L. Barkan Associate Professor Director - Railroad Engineering Program Department of Civil and Environmental Engineering University of Illinois at Urbana-Champaign 1201 Newmark Civil Engineering Laboratory 205 N. Matthews Ave. Urbana, IL, USA Phone: (217) Fax: (217) cbarkan@uiuc.edu Submission Date: 11/15/07 Word Count: 7,750 (5,250 words, 5 Tables, 5 Figures)

2 Schafer & Barkan ABSTRACT Train accident rates are a critical metric of railroad transportation safety and risk performance. Understanding the factors that affect accident rates is also important for evaluating the effectiveness of various accident prevention measures. Accident rates have been the subject of a number of analyses but in general these have not considered the effect of train length on train accident rate. It has been suggested that train accident causes can be classified into two groups, those dependent on train length and correlated with the number of cars in the train, and those independent of train length, corresponding to the number of train-miles operated. These classifications have implications for the quantitative effect of various changes in railroad operating practices on railroad safety performance. Whether an accident cause is a function of car-miles or train miles affects how safety measures that might reduce that cause will affect overall train accident rate. Accident causes have been classified as car or train-mile correlated based on expert opinion but no quantitative test of these classifications has been conducted. The definition of car-mile versus train-mile causes leads to the hypothesis that longer trains should experience more accidents than shorter trains. FRA accident data were used to develop and test a quantitative metric to objectively characterize different accident causes as either car-mile or train-mile correlated. Based on the results of the study a sensitivity analysis was conducted to evaluate how changes in train length affect individual trains' accident rate and system-wide accident rate.

3 Schafer & Barkan Train accident rates are a critical measure of rail transportation safety and risk and understanding them is necessary to evaluate the effect of accident prevention measures. Accident rates have been calculated by various organizations and railroads on a location specific scale and aggregated statistics for all U.S. railroads are published annually by the Federal Railroad Administration (FRA) Office of Safety (1, 2). Rates have been used to assess various factors such as track class, geographic location, train speed, and track type (3-5). However, these analyses have generally not considered the effect of train length on train accident rate. It has been suggested that train length has an effect on accident rate because more cars in a train increase the likelihood that a car or track component may fail and that accident causes can be classified into two types of causes, those that are a function of the number of train-miles operated and those that are a function of car-miles operated (6, 7). The initial classification into these two categories was developed by Arthur D. Little Inc. (ADL) based on the opinions of railroad industry experts. These classifications have implications for the quantitative effect of various changes in practice on railroad safety performance and have been used in subsequent studies of railroad safety (4, 8). Therefore, statistical evaluation of the classifications will enhance their utility and may also clarify our understanding of them. Furthermore, this classification has implications for an accurate understanding of the relationship between train length and accident rate and consequent policy implications for railroad operating practices. We undertook a study to investigate and evaluate the ADL accident cause classifications with the goal of understanding how operating practices, such as train length, affect the likelihood of a train accident. The objectives of this analysis were: Present the methodology for calculating train accident rates based on car-mile and train-mile accident causes, Develop a metric to quantitatively evaluate the classification of accident causes as car or train-mile related, Use the metric to properly classify train accident causes, Provide new train accident rates based on train length using current data, and Conduct a sensitivity analysis on our model to illustrate how changes in train length may affect train accident rate. TRAIN LENGTH BASED ACCIDENT RATES Train accident rates are composed of derailments, collisions, highway-rail grade crossing accidents, and other accident types. The likelihood that a train will be involved in an accident is a function of both car-miles and train-miles operated (7, 9, 10). The number of car-miles operated for a particular train is affected by train length; longer trains accumulate more carmiles. However, not all accident causes are directly related to the length of the train, and instead are related only to the operation of the train. This leads to the concept that train accident causes can be separated into two groups, those dependent on train length, corresponding to the number of car-miles operated, and those independent of train length, corresponding to the number of train-miles operated. They can be defined as follows: "Car-mile-related causes are those for which the likelihood of an accident is proportional to the number of car-miles operated. These include most equipment failures for which accident likelihood is directly proportional to the number of components (e.g. bearing failure) and also include most track component failures for which accident likelihood is proportional to the number of load cycles imposed on the track (e.g. broken rails or welds).

4 Schafer & Barkan Train-mile-related causes are those for which the accident likelihood is proportional to the number of train-miles operated. These include most human error failures for which accident likelihood is independent of train length and depends only on exposure (e.g. grade crossing collisions). (10) Car vs. Train-Mile Expectations The car-mile cause and train-mile cause definitions lead to the hypothesis that longer trains should experience more accidents than shorter trains. This is because longer trains are more susceptible to car-mile-related accidents than shorter trains due to the additional cars in the train. Conversely, a train should experience accidents due to train-mile-related causes regardless of train length. The length of a train, referred to here and throughout the paper, corresponds to the number of cars in the train and not the linear measure of a train s actual length. The hypothesis leads to two predictions that should be evident when examining accident data and can be used to evaluate different train accident causes. The first prediction is that the average length of a train involved in an accident should be greater for car-mile-related causes compared to train-mile-related causes because longer trains will experience a greater proportion of car-mile-related accidents. Conversely, train-mile-related accidents are independent of train length and should not be biased towards long or short trains. The second prediction is that the proportion of accidents for car-mile-related accidents should be an asymptotically increasing function of train length, whereas train-mile-related accidents should be an asymptotically decreasing function. Longer trains should experience a higher percentage of accidents from car-mile-related causes due to their higher percentage of carmiles per train-mile operated. Conversely, shorter trains are expected to experience a greater percentage of accidents from train-mile-related causes. Accident Rate Equation Under the hypothesis that train accidents can be separated into two distinct groups, car-milerelated causes and train-mile-related causes, a new accident rate model that takes into account the two types of classifications can be developed. The new accident rate equation must include a factor for train length to account for accidents that are dependent on the number of car-miles operated. To develop the new model, all FRA train accident causes were examined. The FRA accident database contains 389 unique accident causes (11, 12). A previous study by ADL classified each accident cause as either car-mile or train-mile-related (7). The purpose of this study was to quantify the risk of hazardous material transportation by examining all accident causes. The ADL study showed that accident types should be classified as either car-mile or train-mile-related to properly quantify the car-mile and train-mile related risk. By determining the number of accidents that have occurred due to each cause, two independent and mutually exclusive accident rates can be calculated, the car-mile-accident rate and the train-mile-accident rate. The expected number of accidents that a train will be involved in is the sum of the carmile-accident rate multiplied by number of car-miles and the train-mile-accident rate multiplied by the number of train-miles. The expected number of train accidents that will occur can be calculated as follows:

5 Schafer & Barkan A = R M + R EXP C C T M T where: A EXP = Accidents expected R C = Car-mile-accident rate (accidents per car mile) M C = Number of car miles R T = Train-mile-accident rate (accidents per train mile) M T = Number of train miles Under this model we expect that longer trains will experience more train accidents. As a train s length increases, train-miles operated remains constant, but the number of car-miles increases with each additional car. Therefore, the number of expected accidents for a single train increases due to the additional car-miles (Figure 1a). Number of Accidents Train-Mile-Related TOTAL ACCIDENTS Car-Mile-Related Number of Accidents Train-Mile-Related TOTAL ACCIDENTS Car-Mile-Related Train Length (a) Train Length (b) FIGURE 1 Expected accidents from car-mile and train-mile-related causes as a function of train length for a single train (a) and for a fixed amount of traffic (b). If one extends this model system wide, it suggests the general result that operating longer trains should result in fewer accidents. As train length decreases, more trains are required to move the same number of cars thereby leading to more train-mile-related accidents. Under this simple scenario, accidents will be minimized by running the longest trains feasible given infrastructure and other constraints (Figure 1b). It should be noted that there are limits to the validity of this result for very long train lengths (>150). This is because the hypothesis presented, as well as the data used in our analysis, apply to trains less than this length. In practice it is possible that accident rates for certain trainmile-related accidents may increase as train length becomes very long due to causes such as train handling and train braking. The intention of this analysis is not to suggest that longer trains will necessarily improve safety; instead the purpose is to develop a better quantitative understanding of how changes that affect various accident causes, such as number of trains and train length, will affect overall accident rates.

6 Schafer & Barkan CLASSIFICATION OF ACCIDENT CAUSES To accurately determine the car-mile and train-mile-accident rates, proper classification of each FRA accident cause is needed. The FRA accident cause classification system is very detailed and often includes several variations of one related group of causes. This is a useful attribute of the database, but is more detailed than is necessary for the purpose of this analysis. Consequently, ADL combined similar accident causes into 51 unique groups, 34 of which they classified as car-mile-related (CM) and 17 as train-mile-related (TM) (Table 1) (7). The FRA accident causes are separated into five main groups, mechanical, human, signal, track, and miscellaneous causes. ADL defined most track and mechanical failures as car-mile-related, while most human and signal errors were defined as train-mile-related. The various miscellaneous causes were assigned to either car-mile or train-mile-related. TABLE 1 ADL/AAR Accident Cause Groups and Classification of FRA Accident Causes Group CM/TM Cause Description Group CM/TM Cause Description 01E CM Air Hose Defect (Car) 06H TM Radio Communications Error 02E CM Brake Rigging Defect (Car) 07H TM Switching Rules 03E CM Handbrake Defects (Car) 08H TM Mainline Rules 04E CM UDE (Car or Loco) 09H TM Train Handling (excl. Brakes) 05E CM Other Brake Defect (Car) 10H TM Train Speed 06E CM Centerplate/Carbody Defects (Car) 11H TM Use of Switches 07E CM Coupler Defects (Car) 12H TM Misc. Track and Structure Defects 08E CM Truck Structure Defects (Car) 01M TM Obstructions 09E CM Sidebearing, Suspension Defects (Car) 02M TM Grade Crossing Collisions 10E CM Bearing Failure (Car) 03M CM Lading Problems 11E CM Other Axle/Journal Defects (Car) 04M CM Track-Train Interaction 12E CM Broken Wheels (Car) 05M TM Other Miscellaneous 13E CM Other Wheel Defects (Car) 01S TM Signal Failures 14E CM TOFC/COFC Defects 01T CM Roadbed Defects 15E CM Loco Trucks/Bearings/Wheels 02T TM Non-Traffic, Weather Causes 16E CM Loco Electrical and Fires 03T CM Wide Gauge 17E CM All Other Locomotive Defects 04T CM Track Geometry (excl. Wide Gauge) 18E CM All Other Car Defects 05T CM Buckled Track 19E CM Stiff Truck (Car) 06T CM Rail Defects at Bolted Joint 20E CM Track/Train Interaction (Hunting) (Car) 07T CM Joint Bar Defects 21E CM Current Collection Equipment (Loco) 08T CM Broken Rails or Welds 01H TM Brake Operation (Main Line) 09T CM Other Rail and Joint Defects 02H TM Handbrake Operations 10T CM Turnout Defects-Switches 03H TM Brake Operations (Other) 11T CM Turnout Defects-Frogs 04H TM Employee Physical Condition 12T CM Misc. Track and Structure Defects 05H TM Failure to Obey/Display Signals We used FRA accident data, Rail Equipment Accidents from the FRA Office of Safety, to evaluate the ADL classification of accident causes for the period 1990 to 2005 (11). These data included all accidents occurring on either mainline or siding tracks for all classes of railroads. Accidents on yard and industry tracks were excluded because the average train length for these types of accidents is comparatively shorter due to yard operations. Mainline and siding accidents were combined because of similar accident causes and train length. Car and train-mile relationship predictions for each cause group were compared with the corresponding data from the FRA database. Train lengths were grouped into 10-car bins and the percentage of all carmile-related and train-mile-related accident causes was graphed versus train length (Figure 2).

7 Schafer & Barkan Percent of Total Accidents Residual Error 100% 80% 60% 40% 20% 0% 15% 10% 5% 0% -5% -10% -15% 0-10 y = x R 2 = Car-Mile Train-Mile y = x R 2 = FIGURE 2 Percentage of car and train-mile-related accidents versus train length using the ADL accident cause classification. A power function with residual error is also shown. A regression analysis was conducted in which a power function, of the form y=ax b, was fitted to the data to evaluate how well they conformed to an asymptotically increasing or decreasing functional form. The critical term regarding the curve form of the power function, is the exponent, b. If b > 0, the data are more representative of an asymptotically increasing function (Figure 3a). If b < 0, the data are more representative of an asymptotically decreasing function (Figure 3b). As b approaches zero the power curve becomes less curved and more representative of a horizontal, flat line; whereas for larger absolute values of b, the power function curves more sharply. In the case of b > 0, the function will be convex for b > 1 or concave for b < 1. The residual error from the fitted power curves was also calculated for the various train lengths (Figure 2). The residual error was greatest for long train lengths and trains of less than 10 cars. The results are generally consistent with the car and train-mile predictions. The average length of trains involved in an accident due to car-mile-related causes was 68.3 cars, whereas the average for train-mile-related causes was 52.5 cars. Also, the percentage of train-mile-related accidents declined asymptotically as a function of train length. Although the R 2 values for the regression analysis were significant, it was evident that there were some discrepancies between the observed data and the predicted relationships, as shown in the residual error graph. The error is particularly evident for trains longer than 110 cars Train Length (Number of Cars) >150

8 Schafer & Barkan Percent of Total Accidents (y) b = 0.5 b = 0 b = 1 b = 1.5 Percent of Total Accidents (y) b = 0 b = -0.5 b = -1 b = -1.5 Train Length (x) (a) b FIGURE 3 Characteristics of exponential term, b, of power function y = ax, where (a) represents a car-mile-related cause and (b) represents a train-mile-related cause. These discrepancies suggest that the previous classifications of accident causes by ADL should be evaluated as they may have changed due to the inclusion of this new data and analysis. Therefore, a more detailed analysis of individual accident causes was conducted. The relationships between number of accidents versus train length and percentage of accidents as a function of train length were graphed for each cause group. Although, not all of the accident cause groups contained enough data to allow an accurate evaluation; many of the cause groups conformed well to the predictions for train-mile or car-mile-related causes, examples of which were grade crossing collisions and air hose defects, respectively (Figures 4a and 4b). However, examination of the data also suggested that some of the cause groups need to be reclassified because the results were inconsistent with the car and train-mile predictions (Figures 4c and 4d). A possible explanation exists for the cause group train handling, which is caused by a locomotive engineer improperly handling the train, commonly attributed to excessive horsepower use. ADL defined this as a train-mile-related cause because it is due to human error. However, accidents caused by the use of excessive horsepower are in fact more common in long trains than short trains and therefore resemble a car-mile-related cause. Conversely, the cause group all other locomotive defects was classified by ADL as a car-mile cause because it is a mechanical failure. However, the number of locomotives, and therefore the likelihood of a locomotive defect, is not significantly affected by an increase in cars. Several discrepancies were also observed in other accident cause groups. Therefore a quantitative metric was developed to objectively classify each accident cause group as train-mile or car-mile-related. Development of Classification Metric Train Length (x) We used the two expectations about car and train-mile related causes to develop a quantitative metric to classify each of the ADL accident cause groups. Car-mile accidents should be more prevalent in longer trains and should be an asymptotically increasing function of the percentage of accidents as train length increases, and the reverse should be true for train-mile-related causes. (b)

9 Schafer & Barkan Grade Crossing Collisions (02M) Air Hose Defects (01E) Percent of Accidents 25% 20% 15% 10% 5% 0% Average Train Length = 50.3 y = x R 2 = Train Length (Number of Cars) Percent of Accidents 2.5% 2.0% 1.5% 1.0% 0.5% 0.0% 0-10 Average Train Length = y = 9E-06x R 2 = Train Length (Number of Cars) (a) (b) Percent of Accidents 15% 12% 9% 6% 3% 0% 0-10 Train Handling (excl. Brakes) (09H) Average Train Length = y = x R 2 = Train Length (Number of Cars) (c) All Other Locomotive Defects (17E) y = x R 2 = FIGURE 4 Percentage of accidents versus train length for four example cause groups; correctly classified (a) and (b), and incorrectly classified (c) and (d). Two parameters were calculated for each accident cause to characterize them as either car-mile or train-mile related. The first parameter is the average length of trains involved in an accident for each cause group. The second parameter is derived from the power function curve and its goodness of fit to the data for the percentage of accidents for each cause group as a function of train length. The exponent in the power function was used to assess the asymptotical increase or decrease in the data (Figure 3). The greater the difference between the calculated value of b and zero, the stronger the asymptotically increasing or decreasing function, and therefore the indication of either a car-mile or a train-mile-related cause. For example, cause group 2T, non-traffic/weather causes (b = ), showed a much stronger indication of a train-mile-related cause than 1M, obstructions (b = ). In addition to characterizing the shape of the curves for each accident cause group, it was also important to quantify how well they fit the data. In some cases there were insufficient data to fit a curve and in others the data showed no trend. In order to assess the goodness of fit, the coefficient of determination, R 2, for each data set was calculated. R 2 values range from 0 to 1 and quantify the goodness of fit. Higher values indicate that the curve fits the data better, whereas low values of R 2 indicate a curve that does not. Therefore the lines with a high R 2 are weighted more strongly in the metric than those with a low R 2 value. In summary, the accident metric, which we term AM i, needs to incorporate three characteristics: average length of trains involved in an accident due to a particular accident cause group, the shape of the curve as a function of train length as indicated by the exponent, b, and the goodness of fit of the data to the curve, as indicated by the R 2. The metric is as follows: Percent of Accidents 3.0% 2.4% 1.8% 1.2% 0.6% 0.0% Average Train Length = Train Length (Number of Cars) (d)

10 Schafer & Barkan AM i li = + (bi Ri L 2 ) where: AM i = Accident cause metric for cause group i l i = Average train length for cause group i L = Overall average length of trains involved in accidents in dataset = b i = Value of exponential term in power curve equation, y=ax b, for cause group i R i 2 = Coefficient of Determination for a power curve fit to the data for cause group i If the average length of trains in accidents due to cause i (l i ) is greater than L, AM i is increased and vice versa. The greater the difference between l i and L the more AM i is affected. The second term of the metric is the power function exponent, b. If b i > 0 for cause i it increases AM i ; and vice versa. Similarly, the greater the difference between b i and 0 the greater the effect on AM i. Finally, b is multiplied by R 2 to account for how well the function fits the data. If R 2 is close to 1, the second term will influence the metric more strongly. If the function is a poor fit (low R 2 ), b will have little effect on AM i. Therefore, for R 2 values close to 1 AM i will be calculated equally based on average train length and b; whereas for low R 2 values AM i will be calculated primarily based on average train length. AM i was used to classify and rank the cause groups (Table 2). Not all cause groups included enough data to properly classify them as either car-mile or train-mile-related and these were excluded from the analysis. In particular, cause group 21E, current collection equipment, was excluded because only short passenger trains (<10 cars) were involved in this cause group with none of the accidents resulting in a derailment. The cause groups in Table 2 are ordered from most car-mile-related at the top, to most train-mile-related at the bottom. Cause groups with rankings in the middle are not represented strongly by either car-mile or train-mile classifications. Reclassification of Accident Causes AM i is used to classify accident causes as either more consistent with characteristics of car-milerelated accidents or train-mile-related accidents. If AM i > 1 the cause group is classified as a carmile accident; conversely, if AM i < 1 the cause group is classified as a train-mile-related accident (Table 2). If the classification based on the metric is different from the previous ADL classification this is indicated by a YES in the column heading Change. Using the metric we reclassified 11 cause groups. Cause groups 1H, 9H, and 1S were changed from train-mile to carmile causes. Groups 16E, 17E, 18E, 19E, 1T, 3T, 4T, and 12T were changed from car-mile to train-miles causes. Cause groups 3E, 4E, 14E, 4H, and 11T were not evaluated using the metric due to the small number of accidents for each group. Also, cause group 21E, current collection equipment, was not evaluated because these accidents involved only very short trains that did not typically result in a derailment. The highest ranked car-mile-related accident cause is 1E, air hose defect, with a score of 3.277; whereas the highest ranked train-mile related-accident cause is 02H, handbrake operations, with a score of

11 Schafer & Barkan TABLE 2 Classification, Score, and Rank of Accident Cause Groups Using Metric CAR-MILE-CAUSES Trendline y=ax b Distribution Metric Cause Description a b R 2 Cases Avg. Length Score Rank Change 01E Air Hose Defect (Car) E Broken Wheels (Car) E Bearing Failure (Car) E Other Axle/Journal Defects (Car) H Train Handling (excl. Brakes) YES 01H Brake Operation (Main Line) YES 07E Coupler Defects (Car) E Other Wheel Defects (Car) E Centerplate/Carbody Defects (Car) T Buckled Track E Truck Structure Defects (Car) T Other Rail and Joint Defects M Track-Train Interaction E Other Brake Defect (Car) T Broken Rails or Welds E Brake Rigging Defect (Car) E Track/Train Interaction (Hunting) (Car) T Joint Bar Defects E Sidebearing, Suspension Defects (Car) T Rail Defects at Bolted Joint S Signal Failures YES 10T Turnout Defects-Switches M Lading Problems E Loco Trucks/Bearings/Wheels TRAIN-MILE-CAUSES Trendline y=ax b Distribution Metric Cause Description a b R 2 Cases Avg. Length Score Rank Change 10H Train Speed E Stiff Truck (Car) YES 04T Track Geometry (excl. Wide Gauge) YES 03H Brake Operations (Other) T Roadbed Defects YES 05H Failure to Obey/Display Signals H Use of Switches T Non-Traffic, Weather Causes M Other Miscellaneous E All Other Car Defects YES 12H Misc. Track and Structure Defects T Wide Gauge YES 06H Radio Communications Error E Locomotive Electrical and Fires YES 01M Obstructions M Grade Crossing Collisions E All Other Locomotive Defects YES 07H Switching Rules H Mainline Rules T Misc. Track and Structure Defects YES 02H Handbrake Operations NOT EVALUATED USING METRIC Trendline y=ax b Distribution Metric Cause Description a b R 2 Cases Avg. Length Score Rank Change 04H Employee Physical Condition T Turnout Defects-Frogs E Handbrake Defects (Car) E UDE (Car or Loco) E TOFC/COFC Defects E Current Collection Equipment (Loco)

12 Schafer & Barkan Using the calculated values for AM i we reexamined the overall train-mile and car-milerelated causes for comparison to the ADL classification. Figure 2 indicated that the initial classification was not entirely accurate based on the car and train-mile expectation. After reclassifying the data, the values are now more clearly representative of car-mile and train-milerelated causes (Figure 5). The average train lengths for car-mile-related causes increased from 68.3 to 79.0 cars while the average train length of train-mile-related causes decreased from 52.5 to 48.4 cars. Also, b increased to and R 2 = for car-mile-related causes; whereas, b decreased to and R 2 = for train-mile-related causes. Overall, the new classification is more consistent with the car-mile and train-mile accident predictions. Percent of Total Accidents 100% 80% 60% 40% 20% y = x R 2 = Car-Mile Train-Mile y = x R 2 = % Train Length (Number of Cars) >150 Residual Error 15% 10% 5% 0% -5% -10% -15% FIGURE 5 Percentage of car and train-mile-related accidents versus train length using the new accident cause classification. A power function with residual error is also shown. CALCULATION OF ACCIDENT RATES As stated earlier, train accident rates can be determined by summing the car-mile and train-milerelated rates. The two rates can be calculated using known accident data, the number of car and train-miles operated, and the new classification of accident causes. Data on car-miles and trainmiles operated are available from the AAR (13). Car and train-miles are defined as the

13 Schafer & Barkan movement of a car or train the distance of one mile and is based on the distance run between terminals or stations. Accident information was downloaded and filtered for our criteria from the FRA Office of Safety for the time period (11). FRA data for all accident types for Class I railroads operating on mainline and siding tracks were used to ensure consistency with the AAR definition of car and train-miles for this portion of the analysis. The developed classification metric was used to classify each accident cause. The car and train-mile related accident rates from 1990 to 2005 were calculated by dividing the number of accidents by the number of miles operated (Table 3). In 2005 the accident rate for car-mile-related causes was 1.05x10-8 or about.011 accidents per million car-miles and the train-mile-related accident rate was 8.62x10-7 or about 0.86 accidents per million train-miles. The expected number of train accidents, based on 2005 data, can be calculated as follows: A EXP = 1.05x10 8 M C x10 7 M T where: A EXP = Accidents expected M C = Number of car miles M T = Number of train miles TABLE 3 Car and Train Mainline Accident Rates using the Reclassification of Accident Causes, Class I Freight Railroads, Car-Mile- Caused Accidents Car-Miles Operated (Millions) Car-Mile Accident Rate (per million car miles) Train-Mile- Caused Accidents Train-Miles Operated (Millions) Train-Mile Accident Rate (per million train miles) Year , , , , , , , , , , , , , , , , , , ,913 7, It is clear based on this equation that if the number of cars per train is increased, the consequent increase in car-miles operated leads to an increase in the accident rate for each train so affected. Similarly, an increase in the number of trains operated on a system will increase the number of train-miles operated, and thus increase the number of train-mile-caused accidents. To

14 Schafer & Barkan understand the effect of train length on accident likelihood, the accident rate equation can be expanded to include the term for train length: A 1.05x x10 n d = n d (1.05x x EXP = n d TL TL where: A EXP = Accidents expected n = Number of trains operated d = Number of miles operated T L = Average cars per train (train length) This equation is useful for understanding how changes in operating procedures, such as train length or number of trains operated, will affect the expected number of train accidents. ACCIDENT RATE SENSITIVITY ANALYSIS We conducted two simple analyses of the sensitivity analysis to illustrate the effect of changes in train length on train accident rate. In the first we examine an operational choice of train length given a fixed number of shipments. The analysis parameters are intended to represent a typical high density, long distance, Class I railroad mainline with 25,000 shipments per week and a distance of 2,000 miles with train length and number of trains as the variables. The estimated number of accidents based on 2005 data is 1.05x10-8 accidents per car-mile plus 8.62x10-7 accidents per train mile. We varied train length from 10 cars to 150 cars per train (Table 4). A EXP TABLE 4 Sensitivity Analysis of the Effect of Train Length on Accident Rate Average Train Length (T L ) Number of Trains (n ) Probability of an Accident for each Individual Train Total Expected Number of Accidents 10 2, , ,000 Carloads Shipped; 2,000 Miles; 150 Car Maximum Train Length 8 = n d (1.05x10 TL x10 7 ) )

15 Schafer & Barkan As train length increases, the likelihood that a train will be involved in an accident increases due to the increase in car-miles per train; however, because of the reduction in train miles, the net effect is a reduction in the total number of accidents. So all other things being equal, train accidents will be minimized when train length is maximized or the number of trains operated is minimized. The second study examines how an increase in traffic levels will affect train accident rates. The analysis parameters are similar to those from the previous study of a 2,000 mile Class I railroad freight mainline with the same weekly traffic level of 25,000 shipments. The railroad is currently operating trains with an average length of 100 cars. The shipments are expected to increase by 10% to a new total of 27,500 shipments. The operational choice in this study is either to continue operating the same number, but longer trains, or maintain the current train length and operate more trains. The traffic increase will lead to an increase in overall accidents; however, this effect can be minimized by increasing the length of trains instead of increasing the number of trains operated (Table 5). Again, this study suggests for this type of scenario that a railroad can reduce the overall number of accidents by running fewer, longer trains as opposed to a high number of shorter trains. TABLE 5 Sensitivity Analysis of the Effect of Traffic Increase on Accident Rate Number of Trains (n ) Average Train Length (T L ) Probability of an Accident for each Individual Train Total Expected Number of Accidents ,500 Carloads Shipped; 2,000 Miles CONCLUSIONS Accident rates are affected by both car-mile and train-mile-related accident causes. A consequence of this is that the length of trains affects accident rate. The decision to dispatch the same number of shipments in fewer longer trains versus more, shorter trains will affect the overall accident rate. Furthermore, since some accident causes are correlated with car-miles and others with train-miles, accurate classification of the causes is important to correctly determine the effect of changes on accident rates. The FRA accident causes were combined into 51 unique cause groups, and classified as either car-mile or train-mile related by ADL in A metric was developed to quantitatively evaluate the 51 cause groups based on accident data. Use of the metric led to a reclassification of 11 cause groups. The new classification was found to be more representative of car and train-mile expectations. Mainline car-mile and train-mile-related accident rates were calculated for Class I freight railroads. These rates were used in a sensitivity analysis to illustrate the effect of changes in train length on overall accident rate. Future Work The previous analysis is based on classifications of causes that are either train-mile or car-milerelated. However, many causes may not be purely train or car-mile-related, but instead may depend on a combination of both. Additionally, some causes may depend on both car and train-

16 Schafer & Barkan miles but may be strongly dominated by one or the other. Future work may be possible to define a function for each cause group based on both car-miles and train-miles. Each cause function would weight how strongly the cause is affected by the number of car-miles and the number of train-miles. The developed functions of each cause could then be added together to calculate the effect on overall accident rate. Future work is also possible to examine the affect longer trains may have on different accident types. For example, the operation of longer train lengths may have an effect on the number of grade crossing accidents. Longer trains may lead to fewer incidents of grade crossings collisions due to fewer trains; however, drivers may be more inclined to attempt to pass in front of an oncoming train due to the increased train length and vehicle wait time. It may also be possible to determine an optimal train length to minimize cars derailed. Longer trains may be involved in fewer total accidents, but longer trains may derail or damage more total cars than shorter trains. This is based on the idea that longer trains have more kinetic energy and therefore can derail more cars when involved in an accident. Finally, future work could be completed on comparing the accident model presented in the paper and other accident models. Train accident rates have been developed based on various parameters (3-5). The different train accident rates can be evaluated based on current accident data to test the accuracy of each particular model. It may also be possible to study the combination of different parameters from various accident models to develop a hybrid train accident model. ACKNOWLEDGMENTS The authors appreciate the helpful comments and suggestions from Daniel Peltier and Yung- Cheng Lai, graduate students at the University of Illinois, and Mike Wnek, an undergraduate at the University of Illinois, for his assistance with the analysis. Robert Anderson conducted an initial exploration of the concepts investigated here in connection with his graduate research. The first author was supported by a research grant from the BNSF Railway. REFERENCES 1. Federal Railroad Administration. Railroad Safety Statistics: 2005 Annual Report. Prepared for U.S. Department of Transportation, safetydata.fra.dot.gov/officeofsafety/forms/ Default.asp. Accessed July 31, FRA Office of Safety. Train Accident Trends. safetydata.fra.dot.gov/officeofsafety/query/ Default.asp?page=inctally1.asp. Accessed July 31, TranSys Research, Ltd. Evaluation of Risk Associated with Stationary Dangerous Goods Railway Cars. Prepared for Transport Canada, /14690e.pdf. Accessed July 31, Anderson, R.T. & C.P.L. Barkan. Railroad Accident Rates for Use in Transportation Risk Analysis. In Transportation Research Record: Journal of the Transportation Research Board, No. 1863, TRB, National Research Council, Washington, D.C., 2004, pp Treichel, T.T. & C.P.L. Barkan. Working Paper on Mainline Freight Train Accident Rates. Unpublished Report to the Association of American Railroads, Center for Chemical Process Safety (CCPS). Guidelines for Chemical Transportation Risk Analysis. American Institute of Chemical Engineers, New York, 1995.

17 Schafer & Barkan Arthur D. Little, Inc. Risk Assessment for the Transportation of Hazardous Materials by Rail. Supplementary Report: Railroad Accident Rate and Risk Reduction Option Effectiveness Analysis and Data (2 nd Revision), Surface Transportation Board (STB). Draft Environmental Impact Statement: Finance Docket No San Jacinto Rail Limited and The Burlington Northern and Santa Fe Railway Company Construction and Operation of a Rail Line from the Bayport Loop in Harris County, Texas. Decision ID No , Anderson, R.T. & C.P.L. Barkan. Train Accidents as Functions of Train-Miles and Car- Miles. Working Paper, University of Illinois at Urbana-Champaign, Anderson, R.T. and C.P.L. Barkan Derailment probability analysis and modeling of mainline freight trains. In Proceedings of the 8th International Heavy Haul Conference, Rio de Janeiro, June 2005, pp FRA Office of Safety. Download Data on Demand. safetydata.fra.dot.gov/officeofsafety/ Downloads/Default.asp. Accessed May 10, FRA Office of Safety. Train Accident Cause Codes. safetydata.fra.dot.gov/officeofsafety/ Downloads/Default.asp. Accessed May 10, Association of American Railroads (AAR). Railroad Facts Editions. Association of American Railroads, Washington, D.C.